A Concise Introduction to Decentralized POMDPs by Frans A. Oliehoek, Christopher Amato

By Frans A. Oliehoek, Christopher Amato

This publication introduces multiagent making plans below uncertainty as formalized through decentralized partly observable Markov determination approaches (Dec-POMDPs). The meant viewers is researchers and graduate scholars operating within the fields of man-made intelligence concerning sequential determination making: reinforcement studying, decision-theoretic making plans for unmarried brokers, classical multiagent making plans, decentralized keep an eye on, and operations study.

Show description

Read or Download A Concise Introduction to Decentralized POMDPs PDF

Similar robotics & automation books

Advanced Control for Constrained Processes and Systems

Offers a unified essentially oriented remedy to many limited regulate paradigms-Helps to minimize the space among the on hand limited keep an eye on literature and commercial purposes. Contents: Constraints in suggestions platforms. facing constraints in SISO keep watch over. a few sensible case stories.

Optimal and Robust Estimation: With an Introduction to Stochastic Control Theory, Second Edition

Greater than a decade in the past, world-renowned regulate platforms authority Frank L. Lewis brought what may develop into a regular textbook on estimation, lower than the identify optimum Estimation, utilized in best universities in the course of the global. The time has come for a brand new variation of this vintage textual content, and Lewis enlisted the help of complete specialists to deliver the e-book thoroughly brand new with the estimation equipment riding trendy high-performance structures.

Additional info for A Concise Introduction to Decentralized POMDPs

Sample text

If rewards are to be observed, they should be made part of the observation. 3 Example Domains To illustrate the Dec-POMDP model, we discuss a number of example domains and benchmark problems. These range from the toy (but surprisingly hard) ‘decentralized tiger’ problem to multirobot coordination and communication network optimization. 1 Dec-Tiger We will consider the decentralized tiger (D EC -T IGER) problem Nair et al. [2003c]— a frequently used Dec-POMDP benchmark—as an example. It concerns two agents that are standing in a hallway with two doors.

The core idea is to consider the states and transition, observation and reward functions not as atomic entities, but as consisting of a number of factors, and explicitly representing how different factors affect each other. For instance, in the case of a sensor network, the observations of each sensor typically depend only on its local environment. Therefore, it can be possible to represent the observation model more compactly as a product of smaller observation functions, one for each agent. In addition, since in many cases the sensing costs are local and sensors do not influence their environment there is likely special structure in the reward and transition function.

2004]. This method starts at the last stage, t = h − 1, and works its way back to the first stage, t = 0. As such, we say that DP works backwards through time, or is a bottom-up algorithm. At every stage t the algorithm keeps the solutions that are potentially optimal for the remaining stages t, . . , h − 1. This is similar to dynamic programming for (single-agent) MDP [Puterman, 1994, Sutton and Barto, 1998], but in contrast to that setting it will not be possible to represent these solutions using a simple value function over states.

Download PDF sample

Rated 4.48 of 5 – based on 10 votes